offlinedv2/dreamerv2/agent.py

import tensorflow as tf
import tensorflow_probability as tfp
from tensorflow.keras import mixed_precision as prec

from dreamerv2 import common
from dreamerv2 import expl

tfd = tfp.distributions


class Agent(common.Module):
    def __init__(self, config, obs_space, act_space, step):
        self.config = config
        self.obs_space = obs_space
        self.act_space = act_space['action']
        self.step = step
        self.tfstep = tf.Variable(int(self.step), tf.int64)
        self.wm = WorldModel(config, obs_space, self.tfstep)
        self._task_behavior = ActorCritic(config, self.act_space, self.tfstep)
        if config.expl_behavior == 'greedy':
            self._expl_behavior = self._task_behavior
        else:
            self._expl_behavior = getattr(expl, config.expl_behavior)(
                self.config, self.act_space, self.wm, self.tfstep,
                lambda seq: self.wm.heads['reward'](seq['feat']).mode())
        if self.config.offline_tune_lmbd:
            self.log_lmbd = tf.Variable(0, dtype=tf.float32)
            self.lmbd_optimizer = tf.keras.optimizers.Adam(learning_rate=3e-4)
        else:
            self.lmbd = self.config.offline_lmbd

    @tf.function
    def policy(self, obs, state=None, mode='train'):
        obs = tf.nest.map_structure(tf.tensor, obs)
        tf.py_function(lambda: self.tfstep.assign(
            int(self.step), read_value=False), [], [])
        if state is None:
            latent = self.wm.rssm.initial(len(obs['reward']))
            action = tf.zeros((len(obs['reward']),) + self.act_space.shape)
            state = latent, action
        latent, action = state
        embed = self.wm.encoder(self.wm.preprocess(obs))
        sample = (mode == 'train') or not self.config.eval_state_mean
        latent, _ = self.wm.rssm.obs_step(
            latent, action, embed, obs['is_first'], sample)
        feat = self.wm.rssm.get_feat(latent)
        if mode == 'eval':
            actor = self._task_behavior.actor(feat)
            action = actor.mode()
            noise = self.config.eval_noise
        elif mode == 'explore':
            actor = self._expl_behavior.actor(feat)
            action = actor.sample()
            noise = self.config.expl_noise
        elif mode == 'train':
            actor = self._task_behavior.actor(feat)
            action = actor.sample()
            noise = self.config.expl_noise
        action = common.action_noise(action, noise, self.act_space)
        outputs = {'action': action}
        state = (latent, action)
        return outputs, state

    @tf.function
    def train(self, data, state=None):
        metrics = {}
        state, outputs, mets = self.wm.train(data, state)
        metrics.update(mets)
        start = outputs['post']
        reward = lambda seq: self.wm.heads['reward'](seq['feat']).mode()
        metrics.update(self._task_behavior.train(
            self.wm, start, data['is_terminal'], reward))
        if self.config.expl_behavior != 'greedy':
            mets = self._expl_behavior.train(start, outputs, data)[-1]
            metrics.update({'expl_' + key: value for key, value in mets.items()})
        return state, metrics

    # The folllowing methods split the above for offline training
    @tf.function
    def model_train(self, data):
        _, _, metrics = self.wm.train(data)
        return metrics

    @tf.function
    def agent_train(self, data):
        data = self.wm.preprocess(data)
        embed = self.wm.encoder(data)
        start, _ = self.wm.rssm.observe(embed, data['action'], data['is_first'])

        reward = lambda seq: self.penalized_reward(seq)
        metrics = self._task_behavior.train(self.wm, start, data['is_terminal'], reward)
        if self.config.offline_tune_lmbd:
            metrics['lambda'] = tf.exp(self.log_lmbd)
        else:
            metrics['lambda'] = self.lmbd
        return metrics

    @tf.function
    def compute_penalty(self, seq):
        if self.config.offline_penalty_type == 'log_prob_ens':
            dist = self.wm.rssm.get_dist(seq, ensemble=True)
            penalty = tf.math.reduce_std(dist.log_prob(seq['stoch']), axis=0)
        elif self.config.offline_penalty_type == 'meandis':
            dist = self.wm.rssm.get_dist(seq, ensemble=True)
            m = dist.mean()
            mean_pred = tf.math.reduce_mean(m, axis=0)
            penalty = tf.math.reduce_mean(tf.norm(m - mean_pred, axis=[-1, -2]), axis=0)
        elif self.config.offline_penalty_type == 'mixstd_mean':
            dist = self.wm.rssm.get_dist(seq, ensemble=True)
            gm = tfd.MixtureSameFamily(
                mixture_distribution=tfd.Categorical(probs=[1 / self.config.rssm.ensemble] * self.config.rssm.ensemble),
                components_distribution=tfd.BatchReshape(
                    distribution=dist,
                    batch_shape=dist.batch_shape[1:] + dist.batch_shape[0],
                    validate_args=True
                ))
            penalty = tf.math.reduce_mean(gm.stddev(), axis=[-1, -2])
        else:
            penalty = 0

        return penalty

    @tf.function
    def penalized_reward(self, seq):
        rew = self.wm.heads['reward'](seq['feat']).mode()
        penalty = self.compute_penalty(seq)

        if self.config.offline_tune_lmbd:
            with tf.GradientTape() as tape:
                lmbd = tf.exp(self.log_lmbd)
                lambda_loss = tf.math.reduce_mean(
                    self.log_lmbd * (tf.stop_gradient(lmbd * penalty) - self.config.offline_lmbd_cons))
            variables = [self.log_lmbd]
            grads = tape.gradient(lambda_loss, variables)
            self.lmbd_optimizer.apply_gradients(zip(grads, variables))
        else:
            lmbd = self.lmbd

        return rew - lmbd * penalty

    @tf.function
    def report(self, data):
        report = {}
        data = self.wm.preprocess(data)
        for key in self.wm.heads['decoder'].cnn_keys:
            name = key.replace('/', '_')
            report[f'openl_{name}'] = self.wm.video_pred(data, key)
        return report


class WorldModel(common.Module):
    def __init__(self, config, obs_space, tfstep):
        shapes = {k: tuple(v.shape) for k, v in obs_space.items()}
        self.config = config
        self.tfstep = tfstep
        self.rssm = common.EnsembleRSSM(**config.rssm)
        self.encoder = common.Encoder(shapes, **config.encoder)
        self.heads = {
            'decoder': common.Decoder(shapes, **config.decoder),
            'reward': common.MLP([], **config.reward_head),
        }
        if config.pred_discount:
            self.heads['discount'] = common.MLP([], **config.discount_head)
        for name in config.grad_heads:
            assert name in self.heads, name
        self.model_opt = common.Optimizer('model', **config.model_opt)

    def train(self, data, state=None):
        with tf.GradientTape() as model_tape:
            model_loss, state, outputs, metrics = self.loss(data, state)
        modules = [self.encoder, self.rssm, *self.heads.values()]
        metrics.update(self.model_opt(model_tape, model_loss, modules))
        return state, outputs, metrics

    def loss(self, data, state=None):
        data = self.preprocess(data)
        embed = self.encoder(data)
        post, prior = self.rssm.observe(
            embed, data['action'], data['is_first'], state)
        kl_loss, kl_value = self.rssm.kl_loss(post, prior, **self.config.kl)
        assert len(kl_loss.shape) == 0
        likes = {}
        losses = {'kl': kl_loss}
        feat = self.rssm.get_feat(post)
        for name, head in self.heads.items():
            grad_head = (name in self.config.grad_heads)
            inp = feat if grad_head else tf.stop_gradient(feat)
            out = head(inp)
            dists = out if isinstance(out, dict) else {name: out}
            for key, dist in dists.items():
                like = tf.cast(dist.log_prob(data[key]), tf.float32)
                likes[key] = like
                losses[key] = -like.mean()
        model_loss = sum(
            self.config.loss_scales.get(k, 1.0) * v for k, v in losses.items())
        outs = dict(
            embed=embed, feat=feat, post=post,
            prior=prior, likes=likes, kl=kl_value)
        metrics = {f'{name}_loss': value for name, value in losses.items()}
        metrics['model_kl'] = kl_value.mean()
        metrics['prior_ent'] = self.rssm.get_dist(prior).entropy().mean()
        metrics['post_ent'] = self.rssm.get_dist(post).entropy().mean()
        last_state = {k: v[:, -1] for k, v in post.items()}
        return model_loss, last_state, outs, metrics

    def imagine(self, policy, start, is_terminal, horizon):
        flatten = lambda x: x.reshape([-1] + list(x.shape[2:]))
        start = {k: flatten(v) for k, v in start.items()}
        start['feat'] = self.rssm.get_feat(start)
        start['action'] = tf.zeros_like(policy(start['feat']).mode())
        seq = {k: [v] for k, v in start.items()}
        for _ in range(horizon):
            action = policy(tf.stop_gradient(seq['feat'][-1])).sample()
            state = self.rssm.img_step({k: v[-1] for k, v in seq.items()}, action)
            feat = self.rssm.get_feat(state)
            for key, value in {**state, 'action': action, 'feat': feat}.items():
                seq[key].append(value)
        seq = {k: tf.stack(v, 0) for k, v in seq.items()}
        if 'discount' in self.heads:
            disc = self.heads['discount'](seq['feat']).mean()
            if is_terminal is not None:
                # Override discount prediction for the first step with the true
                # discount factor from the replay buffer.
                true_first = 1.0 - flatten(is_terminal).astype(disc.dtype)
                true_first *= self.config.discount
                disc = tf.concat([true_first[None], disc[1:]], 0)
        else:
            disc = self.config.discount * tf.ones(seq['feat'].shape[:-1])
        seq['discount'] = disc
        # Shift discount factors because they imply whether the following state
        # will be valid, not whether the current state is valid.
        seq['weight'] = tf.math.cumprod(
            tf.concat([tf.ones_like(disc[:1]), disc[:-1]], 0), 0)
        return seq

    @tf.function
    def preprocess(self, obs):
        dtype = prec.global_policy().compute_dtype
        obs = obs.copy()
        for key, value in obs.items():
            if key.startswith('log_'):
                continue
            if value.dtype == tf.int32:
                value = value.astype(dtype)
            if value.dtype == tf.uint8:
                value = value.astype(dtype) / 255.0 - 0.5
            obs[key] = value
        obs['reward'] = {
            'identity': tf.identity,
            'sign': tf.sign,
            'tanh': tf.tanh,
        }[self.config.clip_rewards](obs['reward'])
        if 'discount' not in obs:
            obs['discount'] = 1.0 - obs['is_terminal'].astype(dtype)
        obs['discount'] *= self.config.discount
        return obs

    @tf.function
    def video_pred(self, data, key):
        decoder = self.heads['decoder']
        truth = data[key][:6] + 0.5
        embed = self.encoder(data)
        states, _ = self.rssm.observe(
            embed[:6, :5], data['action'][:6, :5], data['is_first'][:6, :5])
        recon = decoder(self.rssm.get_feat(states))[key].mode()[:6]
        init = {k: v[:, -1] for k, v in states.items()}
        prior = self.rssm.imagine(data['action'][:6, 5:], init)
        openl = decoder(self.rssm.get_feat(prior))[key].mode()
        model = tf.concat([recon[:, :5] + 0.5, openl + 0.5], 1)
        error = (model - truth + 1) / 2
        video = tf.concat([truth, model, error], 2)
        B, T, H, W, C = video.shape
        return video.transpose((1, 2, 0, 3, 4)).reshape((T, H, B * W, C))


class ActorCritic(common.Module):
    def __init__(self, config, act_space, tfstep):
        self.config = config
        self.act_space = act_space
        self.tfstep = tfstep
        discrete = hasattr(act_space, 'n')
        if self.config.actor.dist == 'auto':
            self.config = self.config.update({
                'actor.dist': 'onehot' if discrete else 'trunc_normal'})
        if self.config.actor_grad == 'auto':
            self.config = self.config.update({
                'actor_grad': 'reinforce' if discrete else 'dynamics'})
        self.actor = common.MLP(act_space.shape[0], **self.config.actor)
        self.critic = common.MLP([], **self.config.critic)
        if self.config.slow_target:
            self._target_critic = common.MLP([], **self.config.critic)
            self._updates = tf.Variable(0, tf.int64)
        else:
            self._target_critic = self.critic
        self.actor_opt = common.Optimizer('actor', **self.config.actor_opt)
        self.critic_opt = common.Optimizer('critic', **self.config.critic_opt)
        self.rewnorm = common.StreamNorm(**self.config.reward_norm)

    def train(self, world_model, start, is_terminal, reward_fn):
        metrics = {}
        hor = self.config.imag_horizon
        # The weights are is_terminal flags for the imagination start states.
        # Technically, they should multiply the losses from the second trajectory
        # step onwards, which is the first imagined step. However, we are not
        # training the action that led into the first step anyway, so we can use
        # them to scale the whole sequence.
        with tf.GradientTape() as actor_tape:
            seq = world_model.imagine(self.actor, start, is_terminal, hor)
            reward = reward_fn(seq)
            seq['reward'], mets1 = self.rewnorm(reward)
            mets1 = {f'reward_{k}': v for k, v in mets1.items()}
            target, mets2 = self.target(seq)
            actor_loss, mets3 = self.actor_loss(seq, target)
        with tf.GradientTape() as critic_tape:
            critic_loss, mets4 = self.critic_loss(seq, target)
        metrics.update(self.actor_opt(actor_tape, actor_loss, self.actor))
        metrics.update(self.critic_opt(critic_tape, critic_loss, self.critic))
        metrics.update(**mets1, **mets2, **mets3, **mets4)
        self.update_slow_target()  # Variables exist after first forward pass.
        return metrics

    def actor_loss(self, seq, target):
        # Actions:      0   [a1]  [a2]   a3
        #                  ^  |  ^  |  ^  |
        #                 /   v /   v /   v
        # States:     [z0]->[z1]-> z2 -> z3
        # Targets:     t0   [t1]  [t2]
        # Baselines:  [v0]  [v1]   v2    v3
        # Entropies:        [e1]  [e2]
        # Weights:    [ 1]  [w1]   w2    w3
        # Loss:              l1    l2
        metrics = {}
        # Two states are lost at the end of the trajectory, one for the boostrap
        # value prediction and one because the corresponding action does not lead
        # anywhere anymore. One target is lost at the start of the trajectory
        # because the initial state comes from the replay buffer.
        policy = self.actor(tf.stop_gradient(seq['feat'][:-2]))
        if self.config.actor_grad == 'dynamics':
            objective = target[1:]
        elif self.config.actor_grad == 'reinforce':
            baseline = self._target_critic(seq['feat'][:-2]).mode()
            advantage = tf.stop_gradient(target[1:] - baseline)
            objective = policy.log_prob(seq['action'][1:-1]) * advantage
        elif self.config.actor_grad == 'both':
            baseline = self._target_critic(seq['feat'][:-2]).mode()
            advantage = tf.stop_gradient(target[1:] - baseline)
            objective = policy.log_prob(seq['action'][1:-1]) * advantage
            mix = common.schedule(self.config.actor_grad_mix, self.tfstep)
            objective = mix * target[1:] + (1 - mix) * objective
            metrics['actor_grad_mix'] = mix
        else:
            raise NotImplementedError(self.config.actor_grad)
        ent = policy.entropy()
        ent_scale = common.schedule(self.config.actor_ent, self.tfstep)
        objective += ent_scale * ent
        weight = tf.stop_gradient(seq['weight'])
        actor_loss = -(weight[:-2] * objective).mean()
        metrics['actor_ent'] = ent.mean()
        metrics['actor_ent_scale'] = ent_scale
        return actor_loss, metrics

    def critic_loss(self, seq, target):
        # States:     [z0]  [z1]  [z2]   z3
        # Rewards:    [r0]  [r1]  [r2]   r3
        # Values:     [v0]  [v1]  [v2]   v3
        # Weights:    [ 1]  [w1]  [w2]   w3
        # Targets:    [t0]  [t1]  [t2]
        # Loss:        l0    l1    l2
        dist = self.critic(seq['feat'][:-1])
        target = tf.stop_gradient(target)
        weight = tf.stop_gradient(seq['weight'])
        critic_loss = -(dist.log_prob(target) * weight[:-1]).mean()
        metrics = {'critic': dist.mode().mean()}
        return critic_loss, metrics

    def target(self, seq):
        # States:     [z0]  [z1]  [z2]  [z3]
        # Rewards:    [r0]  [r1]  [r2]   r3
        # Values:     [v0]  [v1]  [v2]  [v3]
        # Discount:   [d0]  [d1]  [d2]   d3
        # Targets:     t0    t1    t2
        reward = tf.cast(seq['reward'], tf.float32)
        disc = tf.cast(seq['discount'], tf.float32)
        value = self._target_critic(seq['feat']).mode()
        # Skipping last time step because it is used for bootstrapping.
        target = common.lambda_return(
            reward[:-1], value[:-1], disc[:-1],
            bootstrap=value[-1],
            lambda_=self.config.discount_lambda,
            axis=0)
        metrics = {'critic_slow': value.mean(), 'critic_target': target.mean()}
        return target, metrics

    def update_slow_target(self):
        if self.config.slow_target:
            if self._updates % self.config.slow_target_update == 0:
                mix = 1.0 if self._updates == 0 else float(
                    self.config.slow_target_fraction)
                for s, d in zip(self.critic.variables, self._target_critic.variables):
                    d.assign(mix * s + (1 - mix) * d)
            self._updates.assign_add(1)